Unsupervised domain adaptation based on cluster matching and Fisher criterion for image classification

Transferring knowledge learned from a labeled domain (source domain) to an unlabeled domain (target domain) is challenging when the two domains have different distributions. The key to the problem is to reduce the distribution shift between the two domains. To align the distributions, most existing works first learn a classifier on the source domain to obtain pseud-labels for target samples, then calculate the target domain distribution based on the pseud-labels. However, the classifier may not meet the target domain because it loses sight of the target distribution during the learning procedure. The mislabeled samples will cause large errors in the calculation of the target domain distribution. To address this issue, we propose a novel method, named cluster matching and Fisher criterion (CMFC), to generate an accurate pseudo-label for each target sample in a latent discriminative subspace by considering both domain distributions. Specifically, we first cluster the samples in both domains, respectively, in the latent subspace and then match the cluster centroid in the target domain with the class centroid in the source domain. Both domain distributions are taken into consideration via cluster matching to assign more accurate pseud-labels. Moreover, we leverage the Fisher criterion to minimize intra-class variances while maximizing inter-class variances, which is conducive to further reducing the distribution shift. We incorporate cluster matching and the Fisher criterion into a united model and design an ADMM algorithm to effectively solve the proposed method. Extensive experiments on five datasets for classification tasks demonstrate the superiority of CMFC.